Statistical research data is the foundation for empirical studies. Researchers in economics or social sciences often obtain such data from external sources through specially designed retrieval interfaces from statistical o ces, commercial data providers as well as from data agencies and other archives. With the advancements in data cataloguing and acquisition of long tail research data sets from individual scientists and institutes, the opportunity is there to install central services for a more holistic data search. In view of a rapid increase in amount of data available and by association an emerging retrieval problem, retrieval interfaces must make e ective use of provided metadata in order to help nd relevant data sets e ciently. This paper presents a multi-step retrieval interface that aims to support the researchers' natural approach to data search and composition. Starting with an idea of the concepts that are to be compared, users kick o their search with thesauri terms and successively specify requirements according to their priorities until suitable data can be selected easily from a manageable number of matching data sets. The prototype presented in this paper also provides means for convenient data harmonization, which is an essential aspect especially when combining statistical data from di erent sources.
A signi cant number of scienti c results are based on research data, since
research has become increasingly data-driven over the years [
With the advancement of computer technology, scienti c analyses are more and more carried out with the aid of machines, as it allows for large amounts of data being processed in short amount of time which has never been possible before. While this certainly is one reason why science has become signi cantly data-driven, it also leads to the fact that most scienti c data is maintained in digital form already. This circumstance and the rise of the Web opens up possibilities for a powerful information infrastructure for supporting these aforementioned goals. Information resources nowadays can be delivered to any place in the world within seconds, laying the ground for delivering the right information to the right place at the right time, the precept of knowledge management.
The Web together with its well-established Web 2.0 technologies has already been recognized as a powerful media for promoting e cient exchange and advancement in the scienti c domain. In this regard, the Leibniz Association has recently started the research alliance Science 2.01 with a growing number of 30 associated institutes to jointly venture into a well-organized and integrated environment of Web-based tools and services for the scienti c community to support rapid exchange and good scienti c practice.
The vision of a thought-out research data infrastructure ts well into this
theme, and many initiatives have formed in the last years, a whole movement to
e ectively enable exchange, citation and preservation of research data. However,
this task has proven non-trivial, as it opened up exhaustive discussions on
metadata schemes2, organized preservation and curation [
At present, e orts are being made to pick up research data as bibliographic
artifacts for re-use, transparency and citation[
In this paper, we investigate how to make use of Semantic Web technologies for providing an e cient and novel approach for the retrieval of statistical data sets that follows a natural approach for data retrieval in the domain of statistics, particularly in the context of economics or the social sciences. Section 2 elaborates on the practice of data acquisition in empirical research to gain a clear picture on the purpose of our system. Related work is discussed in the subsequent section, and Section 4 explains fundamental design decisions and outlines a system architecture. Section 5 describes the user interface itself and how the declared goals have been implemented into features. The paper eventually closes with conclusions and outlook. 1 http://www.leibniz-science20.de 2 particularly important, as in contrast to textual publications, data cannot be understood without documentation 3 http://datahub.io/ 4 http://www.datacite.org/ 2
In many cases, empirical researchers in economics and the social sciences are to put together statistical indicators in large data tables. Typically, each column represents one indicator while the rows represent respective data per year, country or other so-called dimension. The data itself may be self-produced in terms of studies and surveys or acquired from external sources such as statistical o ces, a liated institutes or purchased from commercial data providers. However, common practice is to combine several sources, since some indicators may be obtained from one source while the data for other indicators may be obtained from another one. In this regard, researchers have to be extra careful to make sure respective data represents the same or su ciently similar statistical population.
To gain a clear picture of the goals of this research, we need to clearly
understand the purpose of the system. We have conducted interviews with economic
scientists which helped us gain insights in their work with research data.
Empirical researchers typically start out with an idea of concepts relevant in their
research (e.g. living standards, work conditions, economic growth, etc.). In
addition, they have further details in mind, for instance on reference periods, regions
to be included and distinguished or frequency of data acquisition in case of time
series data. As a result, the data set should be as consistent as possible with
respect to acquisition method, statistical universe and adjustments. To achieve
user acceptance, the system has to be practical in research settings [
As a result, user communication should follow the below steps:
3. Let explore and select matching data sets, allow for revisiting Step 2
After nishing Step 1, data sets associated with the concepts named should be presented to the user. Speci cation of additional requirements should be based on the metadata available for the data sets found. As soon as all relevant requirements are given, the user may inspect and decide on these satisfying data sets and proceed to download at last. 3
There are many repositories on the Web that provide statistical data. Some of them are provided by statistical o ces and data agencies (e.g. Federal Statistical O ce of Germany5, EuroStat6, World Bank7), some are associated with commercial providers (e.g. Thomson Reuters Datastream8, Statista9) and yet others are maintained by journals, archives, libraries or independent organizations (e.g. GESIS10, The Data Hub11, Dataverse repository of Economists Online12). All of these portals are as heterogeneous as the kind and spectrum of data they provide. Some of them provide interfaces for composition of customized data tables where users pick and choose indicators and data records according to their needs. Such features are also provided by the Nesstar system13, one of the most prominent systems for data publishing and online analysis that is being used by a large number of institutes. The Social Science Variables Database at ICPSR14 allows for direct comparison of indicators with respect to a variety of metadata, giving intuitive means to understand di erences in universe, acquisition method and other between data sets. However, users of these systems are to run keyword-based queries and browse through category trees in order to nd relevant data sets individually, and therefore our approach follows a di erent paradigm as presented in Section 2.
Technical challenges in dealing with distributed sources and applying the
OLAP paradigm for retrieval of statistical data from the Linked Data cloud
have been addressed in [
Other approaches are based on semantic links between data sets and research
articles [
Following the steps presented in Section 2, we elaborate on the system architecture of our data retrieval system. To support Step 1, a thesaurus should be used, so that data sets associated with a particular concept can be found easily. To enable the speci cation of requirements, metadata must be given in detail 5 https://www.destatis.de 6 http://epp.eurostat.ec.europa.eu 7 http://data.worldbank.org 8 http://online.thomsonreuters.com/datastream/ 9 http://de.statista.com 10 http://www.gesis.org/en/ 11 http://thedatahub.org 12 http://dvn.iq.harvard.edu/dvn/dv/NEEO 13 http://www.nesstar.com 14 http://www.icpsr.umich.edu and in association with individual indicators and records rather than a separate metadata block for a zipped data bundle. This enables the system to make sense of the data in depth and allow for requirement speci cation as explained later in Section 5.
The research on a data retrieval interface is part of our overall research
activities on an infrastructure for scienti c data for the eld of economics. For
several reasons we regard Semantic Web technologies most suitable for this
purpose, among which is strength in dealing with distributed data and
extensibility, which is required whenever highly speci c long tail data from individual
researchers needs additional vocabulary for description [
The prototype was implemented in Java and JavaScript under the use of the
Play Framework16. The live system was tested on an Apache Tomcat17 and a
Sesame Triple Store18, as the system operates on statistical data provided as
RDF using the RDF Data Cube Vocabulary19 [
The system implements a multi-step retrieval interface as described in Section 2. In the following, we are going to refer to the screenshots given in Figure 1 to 8 in parantheses. Since the expected result is a data table after all, the main screen starts with an empty spreadsheet (1). For Step 1, the user successively enters the names of the concepts that are to be compared in the empty column headers as shown in (2). This task is supported by autocompletion on the basis of concept terms contained in a thesaurus, STW20 in our case. With the selection of a concept, the system displays the number of associated data sets beneath the concept label entered before. A click on this number lists all of them in alphanumerical order (3), and another click reveals a detailed description and further information on the particular data set (7). Yet, at this point, the number of data sets might be huge, and the user may decide to formulate requirements for the data rst as per Step 2. With the selection of a single column header, the panel on the left lists down the union over all properties and property values available in the metadata of all the data sets associated with the concept of the column (4). Hovering over a property or property value produces an info box with documentation on the vocabulary. Selecting a particular property value speci es a requirement and tells the system that only those data sets are relevant for 15 http://www.w3.org/TR/vocab-data-cube/ 16 http://www.playframework.org 17 http://tomcat.apache.org 18 http://www.aduna-software.com/technology/sesame 19 http://www.w3.org/TR/vocab-data-cube/ 20 STW Thesaurus for Economics, http://zbw.eu/stw/ this column that provide this respective property and property value, and the number of relevant data sets drops. With the selection of two or more column headers, the panel on the left shows the intersection between the properties and values of the single columns (5). This feature facilitates harmonization of data, as it reveals which data characteristics can be uni ed among the columns. To specify the contents of the rows, one must specify the Dimension property. A click on the respective header highlights all column headers of the entire table as to indicate that the property of choice must be available in the data sets of all columns. The user selects (multiple) values from the properties listed on the left and the Dimension column lls accordingly (6). This again sets requirements for the data sets, as it lters all data sets that do not provide respective records. Eventually, when all requirements are set, the user examines and selects from the remaining list of data sets for each column (7). If all remaining properties with multiple options are bound to a value, the table lls with actual data content (8). As a last step, the table is o ered for download. Following the call for a research data infrastructure, we have addressed the issue of data retrieval for the domain of economics and social sciences where large amounts of scienti c results are based on statistical data. With the prospect of a rapidly growing amount of data from individual researchers and institutes led in the future, overviewing all relevant data sets e ciently becomes a problem. For this purpose, we have designed an innovative retrieval interface that aims to support researchers in nding and composing data sets according to their natural way of approaching a research question. The prototype presented in this paper provides simple means for data harmonization to enable consistency within statistical population in intuitive ways. Under the use of these features, we expect a signi cant decrease of time needed for data search and composition in comparison to the current practice, although this is yet to be evaluated.
Future improvements of the system should include retrieval from distributed sources, as this version operates on a single triple store endpoint only Moreover, the advantages of using subproperty relations should be investigated and made available to the user. Many other valuable ideas for improvements can be found with regard to user assistance, e.g. warning noti cations when selected time series data include breaks, errors or changes in acquisition method which can be derived from well-maintained metadata.
Finally, this approach needs to be tested on a large archive of various kinds of statistical data and evaluated with end users from the target group of empirical researchers.